首页|改进YOLOv7的遥感图像小目标检测

改进YOLOv7的遥感图像小目标检测

扫码查看
针对在遥感图像中进行小目标检测时,由于目标尺寸小、视觉信息不明显以及背景复杂多变等因素,导致传统的检测方法在精度和鲁棒性方面存在局限性,容易造成漏检、误检的问题,文章提出了一种改进YOLOv7的目标检测算法,采用SIoU作为损失函数,改善目标检测框的定位精度,从而提高检测的准确性和鲁棒性.同时,通过将CNeB模块应用于特征融合过程,增强了特征之间的空间交互能力,进一步提升了小目标的检测性能.此外,为了更好地捕捉遥感图像中小目标的细节特征,利用CA注意力机制,设计了 MPCA模块,实现了对特征图的自适应调整,以提高表征能力.在实验部分,使用了经典的遥感图像数据集进行了大量的实验评估.实验结果表明,所提出的基于SIoU、CNeB模块和CA注意力机制的改进方法在RSOD数据集上四分类均值平均精度达到了 96.8%,比原版YOLOv7提升了 2.5%,有效提高了遥感图像小目标检测精度.
Improved Small Object Detection in Remote Sensing Images Based on YOLOv7
Due to factors such as small target size,obscure visual information and complex and changeable background,the traditional detection methods have limitations in terms of accuracy and robustness,and are prone to miss detection and false detection.In this paper,an improved YOLOv7 target detection algorithm is proposed.SIoU is used as a loss function to improve the positioning accuracy of the target detection frame,so as to improve the accuracy and robustness of the detection.At the same time,by applying CNeB module to the feature fusion process,the spatial interaction between features is enhanced,and the detection performance of small targets is further improved.In addition,in order to better capture the detail features of small and medium-sized targets in remote sensing images,the MPCA module is designed by using the CA attention mechanism to realize the adaptive adjustment of feature maps to improve the representation ability.In the experimental part,a large number of experimental evaluations are carried out using classical remote sensing image dataset.The experimental results show that the proposed method based on SIoU,CNeB module and CA attention mechanism has an average accuracy of 96.8%on the four-class mean on the RSOD dataset,which is 2.5%higher than that of the original YOLOv7,and effectively improves the detection accuracy of small targets in remote sensing images.

target detectionsmall goalYOLOv7loss functionattention mechanism

孙超、周永康、陈正超、刘均学、丁建军

展开 >

江汉大学智能制造学院,武汉 430000

目标检测 小目标 YOLOv7 损失函数 注意力机制

中国国家重点研发计划湖北省教育厅教学研究项目江汉大学校级科研基金资助项目

2018YFD110010420222772022XKZX32

2024

遥感信息
科学技术部国家遥感中心,中国测绘科学研究院

遥感信息

CSTPCD北大核心
影响因子:0.712
ISSN:1000-3177
年,卷(期):2024.39(4)